Two Models for the Embedding of Somatic Mutations Using Drug Screening Datasets

Giovanni Visonà, Gabriele SchweikertZenodo 2020 · DOI

Abstract

Machine learning (ML) models have achieved impressive results when applied to biological and clinical data, but their adaptation to medical practice needs to address serious challenges regarding fairness, interpretability, and explainability. A proper understanding of the mechanistic models underlying cellular biology is crucial to gain insights into the deeper aspects of disease development and response to treatment and is necessary to justify the adoption of the results obtained through ML for clinical applications such as diagnosis and treatment planning. In this report, we use drug screening datasets for the development of methodologies to embed somatic mutations. While clinical data might seem a more obvious candidate for precision medicine applications, drug screens allow us to isolate and analyse the effect of individual factors and their combinations within the cellular model. Insights into the inner workings of these biomarkers within their genomic context have profound implications both for novel treatment development and the understanding of the multi-step carcinogenesis typical of most forms of cancer. We used the Pharmacogenetic Phenome Compendium (PGPC) dataset, which is a drug screen on isogenic cell lines derived from HCT116 colon cancer cells. This data set contains rich phenotypic information with regard to somatic mutations that are highly relevant to the characterisation of the differences between individuals diagnosed with colon cancer. The first approach proposed employs the labels obtained by clustering the phenotypic features for anomaly detection, which flags a combination of drug-mutation as a genetic interaction candidate to validate. Preliminary validation of this method on the PGPC dataset shows that the model can retrieve genetic interactions that are described in the literature. The second method explicitly models the effect of mutation-drug combinations on biological pathways and the effects that these pathways have on the observed features. This model requires additional prior information in the form of pathway membership for genes and drug targets that are collected from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.
BibTeX
@inproceedings{visona2020two,
  author = {Giovanni Visonà and Gabriele Schweikert},
  title = {{Two Models for the Embedding of Somatic Mutations Using Drug Screening Datasets}},
  booktitle = {Zenodo},
  year = {2020},
  doi = {10.5281/zenodo.4018161},
}

Loading related content...

Figure